Multimodal large language models (MLLMs) are increasingly deployed as assistants that interact through text and images, making it crucial to evaluate contextual safety when risk depends on both the visual scene and the evolving dialogue. Existing contextual safety benchmarks are mostly single-turn and often miss how malicious intent can emerge gradually or how the same scene can support both benign and exploitative goals. We introduce the Multi-Turn Multimodal Contextual Safety Benchmark (MTMCS-Bench), a benchmark of realistic images and multi-turn conversations that evaluates contextual safety in MLLMs under two complementary settings, escalation-based risk and context-switch risk. MTMCS-Bench offers paired safe and unsafe dialogues with structured evaluation. It contains over 30 thousand multimodal (image+text) and unimodal (text-only) samples, with metrics that separately measure contextual intent recognition, safety-awareness on unsafe cases, and helpfulness on benign ones. Across eight open-source and seven proprietary MLLMs, we observe persistent trade-offs between contextual safety and utility, with models tending to either miss gradual risks or over-refuse benign dialogues. Finally, we evaluate five current guardrails and find that they mitigate some failures but do not fully resolve multi-turn contextual risks.

Large Language Models (LLMs) have become a popular interface for human-AI interaction, supporting information seeking and task assistance through natural, multi-turn dialogue. To respond to users within multi-turn dialogues, the context-dependent user intent evolves across interactions, requiring contextual interpretation, query reformulation, and dynamic coordination between retrieval and generation. Existing studies usually follow static rewrite, retrieve, and generate pipelines, which optimize different procedures separately and overlook the mixed-initiative action optimization simultaneously. Although the recent developments in deep search agents demonstrate the effectiveness in jointly optimizing retrieval and generation via reasoning, these approaches focus on single-turn scenarios, which might lack the ability to handle multi-turn interactions. We introduce a conversational agent that interleaves search and reasoning across turns, enabling exploratory and adaptive behaviors learned through reinforcement learning (RL) training with tailored rewards towards evolving user goals. The experimental results across four widely used conversational benchmarks demonstrate the effectiveness of our methods by surpassing several existing strong baselines.

Advancing beyond single monolithic language models (LMs), recent research increasingly recognizes the importance of model collaboration, where multiple LMs collaborate, compose, and complement each other. Existing research on this topic has mostly been disparate and disconnected, from different research communities, and lacks rigorous comparison. To consolidate existing research and establish model collaboration as a school of thought, we present MoCo: a one-stop Python library of executing, benchmarking, and comparing model collaboration algorithms at scale. MoCo features 26 model collaboration methods, spanning diverse levels of cross-model information exchange such as routing, text, logit, and model parameters. MoCo integrates 25 evaluation datasets spanning reasoning, QA, code, safety, and more, while users could flexibly bring their own data. Extensive experiments with MoCo demonstrate that most collaboration strategies outperform models without collaboration in 61.0

Activation steering controls language model behavior by adding directions to internal representations at inference time, but standard residual-stream steering can fail in stateful dialogue. We identify KV-cache contamination as a key failure mode: steered token states are stored and repeatedly reused, turning a local perturbation into cumulative coherence degradation. To address this challenge, we propose Gated Cropped Attention-Delta steering (GCAD), which extracts steering signals from system-prompt contributions to self-attention and applies them with token-level gating. Across persona-steering experiments, GCAD preserves trait control while substantially improving long-horizon coherence. On the main multi-turn benchmark, GCAD improves average coherence drift from -18.6 to -1.9 and raises turn-10 trait expression from 78.0 to 93.1. These results suggest that activation steering becomes more reliable when interventions follow the prompt-mediated pathways that models already use for behavioral control.

Personalized large language models (LLMs) tailor content to individual preferences using user profiles or histories. However, existing parameter-efficient fine-tuning (PEFT) methods, such as the “One-PEFT-Per-User” (OPPU) paradigm, require training a separate adapter for each user, making them computationally expensive and impractical for real-time updates. We introduce Profile-to-PEFT, a scalable framework that employs a hypernetwork, trained end-to-end, to map a user's encoded profile directly to a full set of adapter parameters (e.g., LoRA), eliminating per-user training at deployment. This design enables instant adaptation, generalization to unseen users, and privacy-preserving local deployment. Experimental results demonstrate that our method outperforms both prompt-based personalization and OPPU while using substantially fewer computational resources at deployment. The framework exhibits strong generalization to out-of-distribution users and maintains robustness across varying user activity levels and different embedding backbones. The proposed Profile-to-PEFT framework enables efficient, scalable, and adaptive LLM personalization suitable for large-scale applications.

Real-world large language model deployments (e.g., conversational AI systems, code generation assistants) naturally generate abundant implicit user dissatisfaction (DSAT) signals, as users iterate toward better answers through refinements, corrections, and expressed preferences, while explicit satisfaction (SAT) feedback is scarce. Existing preference learning approaches are poorly aligned with this data profile, as they rely on costly human annotations or assume plentiful positive responses. In this paper, we introduce \textbf{DRIFT} (\textbf{D}issatisfaction-\textbf{R}efined \textbf{I}terative pre\textbf{F}erence \textbf{T}raining), which anchors training on real-world DSAT signals and samples positives dynamically from the evolving policy. Empirically, DRIFT models trained on real-world \textit{WildFeedback} datasets and synthetic \textit{UltraFeedback} datasets achieve up to +6.23\% (7B) / +7.61\% (14B) on WildBench Task Score and up to +8.95\% (7B) / +12.29\% (14B) on AlpacaEval2 win rate over base models, outperforming strong baseline methods such as iterative DPO and SPIN. At larger scales, the improvements are particularly pronounced: 14B models trained with DRIFT surpass GPT-4o-mini on WildBench. Further analysis shows that DRIFT also preserves exploratory capacity, yielding more diverse high-reward solutions rather than collapsing to narrow subsets. Theoretically, we demonstrate that this design preserves preference margins and avoids the gradient degeneration. These results show that DRIFT is an effective and scalable recipe for real-world post-training that leverages the most abundant and informative signal.

Personalized systems rely on user representations to connect behavioral history with downstream recommendation applications. Existing methods typically employ either supervised latent user embeddings, which are effective for retrieval but difficult to interpret, or textual user profiles, which are interpretable but challenging to optimize for downstream utility due to lack of direct supervision. To bridge this gap, we present BLUE, a reinforcement learning framework that unifies these two forms of user representation by aligning language-based user profiles with embedding-based recommendation objectives. Given a user interaction history, BLUE leverages a profiler Large Language Model (LLM) to generate textual profiles, while an embedding model provides reward signals. This encourages the resulting textual representations to move closer to positive items and farther from negative ones in the embedding space. We further introduce a text-space supervision signal based on next-item prediction, ensuring the learned profiles remain both semantically meaningful and highly effective for downstream retrieval. Experiments on Amazon Reviews 2023 and Google Local Reviews in zero-shot sequential recommendation settings demonstrate that BLUE consistently outperforms strong baselines under both frozen and trainable embedding conditions. Notably, BLUE achieves clear gains in cross-domain transfer, highlighting the strong generalization ability of the learned user profiles. Furthermore, these generated profiles provide superior personalized context for question answering compared to raw user histories or alternative profile optimization methods. Overall, these results show that BLUE provides an effective way to unify interpretable textual profiling with discriminative latent embeddings for personalization.

Previous research has sought to enhance the graph reasoning capabilities of LLMs by supervised fine-tuning on synthetic graph data. While these led to specialized LLMs better at solving graph algorithm problems, we don't need LLMs for shortest path: we need generalization from synthetic graph data to real-world tasks with implicit graph structures. In this work, we propose to unlock generalizable learning of graph with post-training alignment with synthetic data. We first design solution-based and process-based rewards for synthetic graph problems: instead of rigid memorizing response patterns in direct fine-tuning, we posit that post-training alignment would help LLMs grasp the essentials underlying graph reasoning and alleviate overfitting on synthetic data. We employ post-training alignment algorithms such as GRPO and DPO, aligning both off-the-shelf LLMs and LLMs fine-tuned on synthetic graph data. We then compare them against existing settings on both in-domain synthetic tasks and out-of-domain real-world tasks with implicit graph structures such as multi-hop QA, structured planning, and more. Extensive experiments demonstrate that our post-training alignment recipe leads to statistically significant improvement on 5 datasets, with an average gain of 12.9

Taxonomies play a crucial role in various applications by providing a structural representation of knowledge. The task of taxonomy expansion involves integrating emerging concepts into existing taxonomies by identifying appropriate parent concepts for these new query concepts. Previous approaches typically relied on self-supervised methods that generate annotation data from existing taxonomies. However, these methods are less effective when the existing taxonomy is small (fewer than 100 entities). In this work, we introduce CodeTaxo, a novel approach that leverages large language models through code language prompts to capture the taxonomic structure. Extensive experiments on five real-world benchmarks from different domains demonstrate that CodeTaxo consistently achieves superior performance across all evaluation metrics, significantly outperforming previous state-of-the-art methods. The code and data are available at .

The instruction hierarchy, which establishes a priority order from system messages to user messages, conversation history, and tool outputs, is essential for ensuring consistent and safe behavior in language models (LMs). Despite its importance, this topic receives limited attention, and there is a lack of comprehensive benchmarks for evaluating models' ability to follow the instruction hierarchy. We bridge this gap by introducing IHEval, a novel benchmark comprising 3,538 examples across nine tasks, covering cases where instructions in different priorities either align or conflict. Our evaluation of popular LMs highlights their struggle to recognize instruction priorities. All evaluated models experience a sharp performance decline when facing conflicting instructions, compared to their original instruction-following performance. Moreover, the most competitive open-source model only achieves 48 accuracy in resolving such conflicts. Our results underscore the need for targeted optimization in the future development of LMs.

Generative models such as Large Language Models (LLM) and Multimodal Large Language models (MLLMs) trained on massive web corpora can memorize and disclose individuals' confidential and private data, raising legal and ethical concerns. While many previous works have addressed this issue in LLM via machine unlearning, it remains largely unexplored for MLLMs. To tackle this challenge, we introduce Multimodal Large Language Model Unlearning Benchmark (MLLMU-Bench), a novel benchmark aimed at advancing the understanding of multimodal machine unlearning. MLLMU-Bench consists of 500 fictitious profiles and 153 profiles for public celebrities, each profile feature over 14 customized question-answer pairs, evaluated from both multimodal (image+text) and unimodal (text) perspectives. The benchmark is divided into four sets to assess unlearning algorithms in terms of efficacy, generalizability, and model utility. Finally, we provide baseline results using existing generative model unlearning algorithms. Surprisingly, our experiments show that unimodal unlearning algorithms excel in generation and cloze tasks, while multimodal unlearning approaches perform better in classification tasks with multimodal inputs.

Personalized conversational information retrieval (CIR) systems aim to satisfy users' complex information needs through multi-turn interactions by considering user profiles. However, not all search queries require personalization. The challenge lies in appropriately incorporating personalization elements into search when needed. Most existing studies implicitly incorporate users' personal information and conversational context using large language models without distinguishing the specific requirements for each query turn. Such a “one-size-fits-all” personalization strategy might lead to sub-optimal results. In this paper, we propose an adaptive personalization method, in which we first identify the required personalization level for a query and integrate personalized queries with other query reformulations to produce various enhanced queries. Then, we design a personalization-aware ranking fusion approach to assign fusion weights dynamically to different reformulated queries, depending on the required personalization level. The proposed adaptive personalized conversational information retrieval framework APCIR is evaluated on two TREC iKAT datasets. The results confirm the effectiveness of adaptive personalization of APCIR by outperforming state-of-the-art methods.

Best-of-N decoding methods instruct large language models (LLMs) to generate multiple solutions, score each using a scoring function, and select the highest scored as the final answer to mathematical reasoning problems. However, this repeated independent process often leads to the same mistakes, making the selected solution still incorrect. We propose a novel prompting method named Stepwise Correction (StepCo) that helps LLMs identify and revise incorrect steps in their generated reasoning paths. It iterates verification and revision phases that employ a process-supervised verifier. The verify-then-revise process not only improves answer correctness but also reduces token consumption with fewer paths needed to generate. With StepCo, a series of LLMs demonstrate exceptional performance. Notably, using GPT-4o as the backend LLM, StepCo achieves an average accuracy of 94.1 across eight datasets, significantly outperforming the state-of-the-art Best-of-N method by +2.4, while reducing token consumption by 77.8%. Our implementation is made publicly available at https://wzy6642.github.io/stepco.github.io.

Large language model (LLM) personalization aims to tailor model behavior to individual users based on their historical interactions. However, its effectiveness is often hindered by two key challenges: the cold-start problem, where users with limited history provide insufficient context for accurate personalization, and the biasing problem, where users with abundant but skewed history cause the model to overfit to narrow preferences. We identify both issues as symptoms of a common underlying limitation, i.e., the inability to model collective knowledge across users. To address this, we propose a local-global memory framework (LoGo) that combines the personalized local memory with a collective global memory that captures shared interests across the population. To reconcile discrepancies between these two memory sources, we introduce a mediator module designed to resolve conflicts between local and global signals. Extensive experiments on multiple benchmarks demonstrate that LoGo consistently improves personalization quality by both warming up cold-start users and mitigating biased predictions. These results highlight the importance of incorporating collective knowledge to enhance LLM personalization.

Generative models such as Large Language Models (LLMs) and Multimodal Large Language Models (MLLMs) trained on massive datasets can lead them to memorize and inadvertently reveal sensitive information, raising ethical and privacy concerns. While some prior works have explored this issue in the context of LLMs, it presents a unique challenge for MLLMs due to the entangled nature of knowledge across modalities, making comprehensive unlearning more difficult. To address this challenge, we propose Modality Aware Neuron Unlearning (MANU), a novel unlearning framework for MLLMs designed to selectively clip neurons based on their relative importance to the targeted forget data, curated for different modalities. Specifically, MANU consists of two stages: important neuron selection and selective pruning. The first stage identifies and collects the most influential neurons across modalities relative to the targeted forget knowledge, while the second stage is dedicated to pruning those selected neurons. MANU effectively isolates and removes the neurons that contribute most to the forget data within each modality, while preserving the integrity of retained knowledge. Our experiments conducted across various MLLM architectures illustrate that MANU can achieve a more balanced and comprehensive unlearning in each modality without largely affecting the overall model utility.

Learning from preference feedback is essential for aligning large language models (LLMs) with human values and improving the quality of generated responses. However, existing preference learning methods rely heavily on curated data from humans or advanced LLMs, which is costly and difficult to scale. In this work, we present PUGC, a novel framework that leverages implicit human Preferences in unlabeled User-Generated Content (UGC) to generate preference data. Although UGC is not explicitly created to guide LLMs in generating human-preferred responses, it often reflects valuable insights and implicit preferences from its creators that has the potential to address readers’ questions. PUGC transforms UGC into user queries and generates responses from the policy model. The UGC is then leveraged as a reference text for response scoring, aligning the model with these implicit preferences. This approach improves the quality of preference data while enabling scalable, domain-specific alignment. Experimental results on Alpaca Eval 2 show that models trained with DPO and PUGC achieve a 9.37% performance improvement over traditional methods, setting a 35.93% state-of-the-art length-controlled win rate using Mistral-7B-Instruct. Further studies highlight gains in reward quality, domain-specific alignment effectiveness, robustness against UGC quality, and theory of mind capabilities. Our code and dataset are available at https://zhaoxuan.info/PUGC.github.io/.

Multimodal Large Language Models (MLLMs) are increasingly deployed in real-world applications, yet their ability to make context-aware safety decisions remains limited. Existing methods often fail to balance oversensitivity (unjustified refusals of benign queries) and undersensitivity (missed detection of visually grounded risks), leaving a persistent gap in safety alignment. To address this issue, we introduce Safety-aware Contrastive Decoding (SafeCoDe), a lightweight and model-agnostic decoding framework that dynamically adjusts token generation based on multimodal context. SafeCoDe operates in two stages: (1) a contrastive decoding mechanism that highlights tokens sensitive to visual context by contrasting real and Gaussian-noised images, and (2) a global-aware token modulation strategy that integrates scene-level reasoning with token-level adjustment to adapt refusals according to the predicted safety verdict. Extensive experiments across diverse MLLM architectures and safety benchmarks, covering undersensitivity, oversensitivity, and general safety evaluations, show that SafeCoDe consistently improves context-sensitive refusal behaviors while preserving model helpfulness.

Large Language Models (LLMs) excel in various tasks, including personalized recommendations. Existing evaluation methods often focus on rating prediction, relying on regression errors between actual and predicted ratings. However, user rating bias and item quality, two influential factors behind rating scores, can obscure personal preferences in user-item pair data. To address this, we introduce PerRecBench, disassociating the evaluation from these two factors and assessing recommendation techniques on capturing the personal preferences in a grouped ranking manner. We find that the LLM-based recommendation techniques that are generally good at rating prediction fail to identify users' favored and disfavored items when the user rating bias and item quality are eliminated by grouping users. With PerRecBench and 19 LLMs, we find that while larger models generally outperform smaller ones, they still struggle with personalized recommendation. Our findings reveal the superiority of pairwise and listwise ranking approaches over pointwise ranking, PerRecBench's low correlation with traditional regression metrics, the importance of user profiles, and the role of pretraining data distributions. We further explore three supervised fine-tuning strategies, finding that merging weights from single-format training is promising but improving LLMs' understanding of user preferences remains an open research problem. Code and data are available at https://github.com/TamSiuhin/PerRecBench

The scaling laws have become the de facto guidelines for designing large language models (LLMs), but they were studied under the assumption of unlimited computing resources for both training and inference. As LLMs are increasingly used as personalized intelligent assistants, their customization (i.e., learning through fine-tuning) and deployment onto resource-constrained edge devices will become more and more prevalent. An urgent but open question is how a resource-constrained computing environment would affect the design choices for a personalized LLM. We study this problem empirically in this work. In particular, we consider the tradeoffs among a number of key design factors and their intertwined impacts on learning efficiency and accuracy. The factors include the learning methods for LLM customization, the amount of personalized data used for learning customization, the types and sizes of LLMs, the compression methods of LLMs, the amount of time afforded to learn, and the difficulty levels of the target use cases. Through extensive experimentation and benchmarking, we draw a number of surprisingly insightful guidelines for deploying LLMs onto resource-constrained devices. For example, an optimal choice between parameter learning and RAG may vary depending on the difficulty of the downstream task, the longer fine-tuning time does not necessarily help the model, and a compressed LLM may be a better choice than an uncompressed LLM to learn from limited personalized data.